Elevator Control Strategies

Bachelors essay in computer science

Royal Institute of Technology

Johanna Axelsson
Kungshamra 3, 170 70 Solna
0736-574676
joaxel@kth.se
Sarah Bernelind
Ellen Keys gata 33, 129 52 Hagersten
0702-420636
sarahber@kth.se

Supervisor: Johan Boye

May 29, 2013

Abstract
The purpose of this essay is to investigate if it is effective to switch strategies
for elevators during one day in an office building. This essay describes some
of the strategies in use today, followed by a comparison and analysis of two
of the strategies described. We have also implemented optimizations to one
of these strategies. From our test results we can conclude that our optimized
strategy worked and produced better results on average waiting time and
total traveling time than the two strategies we previously implemented and
that it is not necessarily effective to completely change strategies but change
parts of them depending on the traffic flow.

Referat
Syftet med denna rapport ar att undersoka om det ar effektivt att byta
strategi for hissar i en kontorsbyggnad under en dags olika trafikfloden. Forst
beskrivs n
agra av de vanligaste strategierna, vilket foljs av en jamforelse
och analys av tv
a av dessa. For att kunna jamfora dessa strategier har en
implementation av ett program som simulerar en byggnad med hissar gjorts.
I programmet anvands sedan dessa strategier for att ta fram testresultat. Den
mest effektiva av dessa tv
a strategier har optimerats och jamforts ytterligare
med de tidigare resultaten. Testresultaten visar att den optimerade strategin
var den mest effektiva med avseende p
a genomsnittsvantetiden och den totala
genomsnittsrestiden, och att det inte ar nodvandigt att helt byta strategi
under dagen, men att det lonar sig att optimera delar av den, beroende p
a
trafikflodet.

Preface
This essay and the corresponding project have been carried out as part of a
(15 credits) bachelor degree at the CSC department of the Royal Institute of
Technology.
The background research was done by Sarah, and the implementation by
Johanna. Both authors participated in the writing of results, discussion and
conclusion.
We want to thank our supervisor Johan Boye for help and guidance during
this project.

Contents
1 Introduction
1
Problem Statement . . . . . . . . . . . . . . . . . . . . . . . .

2
2

2 Background
1
Elevator Control . . . . . . . . . .
2
Development of Strategies . . . . .
2.1
Collective Control Strategy
2.2
Zone Approach Strategy . .
2.3
Search-based Strategy . . .
2.4
Rule-based Strategy . . . .
2.5
Genetic Algorithms . . . . .
3
Passenger Arrival . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

3
3
3
4
4
4
5
6
6

3 Method
1
Elevator Simulator . . .
2
Development of the First
3
Simulation and Statistics
4
Optimized strategies . .
4.1
First Version . .
4.2
Second Version .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

8
8
11
11
12
12
12

.
.
.
.
.

13
13
13
15
15
16

4 Results
1
Initial Strategies
2
Optimization . .
3
Higher Pressure .
4
Energy Efficiency
5
Source of Error .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

. . . . . .
Strategies
. . . . . .
. . . . . .
. . . . . .
. . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

5 Discussion

18

6 Conclusion

20

References

21

Chapter 1
Introduction
Today, elevator group control is used worldwide and there are many different
ways of deciding which elevator should handle a request. During a day in an
office building, an elevator group control system will confront many different
flows of passengers. In the morning, there is an up-peak in the traffic, during
the day the traffic is mixed among the floors and in the afternoon, the flow
out of the building is heavy.
This report investigates how effective different elevator group control
strategies are on a system with two elevators, and which optimizations that
can improve the performance of these strategies. To be able to compare the
strategies, we have written a program which simulates a building with two
elevators, and passengers flowing through the system. The results are then
compared to see which control strategy fits the situation best. From studying
this data, we developed our own optimized strategy from the most successful
one to see what difference these improvements make.
Since the traffic through a building differs much during a day it may
be efficient to change strategy during the day, adjusting it to the flow of
passengers. Ultimately, we will discover which strategy fits what flow best.

Problem Statement

We will compare different elevator control strategies with respect to the average waiting time and total travelling time in an office building with two
elevators. These strategies will be compared during two different stages of the
day; the up-peak in the morning when passengers arrive from an lobby floor
and the down-peak when passengers are leaving the building. Thus there
will be two different flows and the question is whether it is more efficient to
change strategies during these flows.

Chapter 2
Background
1

Elevator Control

Most of the research done today on elevator group control deals with systems
where there is more than one car involved. Most engineers tend to measure
the performance of the system in times such as average waiting time; the
time between the passenger presses the call button until the elevator arrives,
and average service time; time on average from which the passenger presses
the button until it has reached its destination[1].
Another aspect of elevator movements is the energy consumption of the
system. In fact, 3-5 percent of the total energy consumption in a building
comes from lifts, escalators and moving walks, according to the E4 project
from the European Commission[2]. Therefore, minimizing the travel distance
for the elevator is important.
In the last decades new techniques have been introduced, for example, fuzzy logic, neural networks and evolutionary algorithms (genetic algorithms), which we will describe more thoroughly later in the essay. Today,
hybrid techniques using the best of these methods have made the problem
of engineering an effective elevator control system an area of computational
intelligence[3][4], an area of which the focus is on problems normaly solved
by an actual brain.

Development of Strategies

Many strategies for controlling elevators have been developed during the
years. The first ones were very simple, straightforward strategies, but today it
is common to see algorithms using artificial intelligence and machine learning
to improve the elevators movements. We have researched some of the more
interesting strategies and algorithms to see what kind of strategies exist.

CHAPTER 2. BACKGROUND

2.1

Collective Control Strategy

Among the oldest strategies is the collective control strategy. It is a simple

strategy for an elevator. The elevator runs in one direction and picks up
passengers with the same direction as the elevator. When all of the passengers
have been dropped off and there are no more requests in that direction, the
elevator will go in the other direction if there are any requests. Else it will
just stop and be idle where it dropped the last passenger[5]. One drawback of
this strategy is the phenomenon called bunching, where several cars answer
the same call from a floor and arrive at similar times, thus increasing both
the waiting time for the other passengers in the system and the elevators
travel distance.

2.2

Zone Approach Strategy

In the zone approach strategy, which is based on the collective control strategy, each elevator is assigned a zone of the building. The car only answers
calls from that area and stops in the zone when it is idle. However, it is
allowed for the car to drop off passengers outside its zone. This strategy
aims to keep the cars separated and to avoid several cars answering the same
call. It is suited for heavy traffic when the hall calls are spread all through
the building but it loses, at the same time, a lot of flexibility[6] since the cars
cannot cover for each other.
To create an optimal strategy the distribution of zones has to be chosen
carefully. When making the decision on how to divide the building there
might be several variables to consider. Zones might be divide depending on
the population of the floors or if there is a floor of importance, such as an
executive floor. The general notion is to have as many zones as you have
elevators[7].

2.3

Search-based Strategy

Unlike the algorithms described above, a search-based strategy is based on

a search algorithm rather than a greedy variant which would chose the first
solution, which is not always the best solution. It will search through the
possibilities to assign, based on some criterion, a car to a call. The search
space consists of all the hall calls made in the building matched with the
elevators and the criterion on which you search could be minimizing the
4

CHAPTER 2. BACKGROUND

average traveling time or the average waiting time. These types of algorithms
are used when you want to find a minimum value of a mathematical equation
or find an item with specific traits among a specific collection of items.
By using a greedy algorithm you can shorten the time for assigning the
call to a car since the greedy algorithm immediately assigns a car based on
the currently available data. This is good for minimizing the average waiting
time but it is not flexible because the greedy algorithm never reassesses its
choice of call assignment, it base its decision localy, meaning it will make
the best decision from the point where it stands at that exact moment. The
opposite of a greedy algorithm is a non-greedy algorithm which is flexible and
can reassess its call assignments in the light of new continuous information
from the elevator system. The non-greedy algorithm will take more time to
decide which call it will assign and thus the average waiting time might be
longer but the overall results might be better[6].

2.4

Rule-based Strategy

The rule-based strategy is based on IF condition -THEN action logic. The

strategy is commonly used in Artificial Intelligence but can in some sense be
applied to all control strategies. A rule base is created out of expertise and
research, and from there the elevator control makes its decisions. When this
type of strategy is used you can combine it with fuzzy logic. Fuzzy logic is a
form of many-valued logic. In fuzzy logic the concept is to use variables that
can be more than just boolean values, for example, a temperature variable
could be warm, cold or slightly above freezing. All of these different
values on the temperature variable have different truth values.[8]. Utilizing
this opens a whole new dimension of decisions needing to be made by the
system. For example, one rule could be IF(there is a call from a lower floor)
AND (all the cars are descending) THEN (assign the descending car that will
have the best results based on the average waiting time in the system)[9].
When you look at our example the call variable is set to lower floor
and the cars variable is set to descending. It is by evaluating these two
variables that you get your THEN action. You would get a different THEN
action if the two variables where different.

CHAPTER 2. BACKGROUND

2.5

Genetic Algorithms

Some problems can also be solved using a genetic algorithm, which imitates
the process of evolution. [10]. Genes consists of chromosomes and we inherit
these genes from our parents. In order to artificially mimic the process of
natural selection the chromosomes to start with has to be created. When
programming genetic algorithms these chromosomes are designed and, put
together rendomly as a gene or individual, poses a solution to the problem
at hand. The genes are being tested to decide their fitness score, which is a
number indicating how successful the gene is[11]. When selecting genes for
mating, the fitness score is used to decide which genes should be selected.
Because the selection of the chromosomes is weighted, the probability is high
that two genes with a high fitness score are selected. Thus the population
will evolve to the better in each iteration, just as it would have for living
creatures.
When applied to the problem of engineering an effective elevator control
system the chromosomes put together as a gene would describe how the system would act in different situations and the chromosomes with the result
nearest the demand of the engineers would be awarded the highest fitness
score. By mating genes in each iteration the results will only get better
and better after the bad ones are being sieved out. This trait leads to the advantage that you, at any moment, can stop it and still have a better solution
than the one you started with. The disadvantage with a genetic algorithm
is that it is iterative and therefore slow in execution.
To summarize, there are a lot of interesting strategies you can use when
you want to create an elevator control system and many of them are being
used today. Which one is the best depends on what conditions you have
but genetic algorithms have become popular in the last years[11] and a lot
of research is being done on the subject which is being used in all kinds of
different areas.

Passenger Arrival

During a day, the rate in which passengers arrive to a building varies and it
is the passenger arrival that impels an elevator system. In this report, the
focus is set on an office building and its typical flow.

CHAPTER 2. BACKGROUND

Figure 2.1: Traffic flows through a building in up-peak and down-peak

In up-peak traffic there will be one arrival floor and many destination
floors while in down-peak there will be many arrival floors and one destination
floor, as shown in Figure 2.1. In heavy up-peak traffic, the cars will fill up
and stop at many floors thus taking a longer time to make the trip and the
average waiting time will be longer for the passengers waiting in the lobby.
In heavy down-peak traffic the cars might fill up on the upper floors thus
making the total round-trip time for the car less than for the same car in
heavy up-peak traffic.
The situation will be reversed when there is light ascending traffic. By
then cars can be idle in the lobby which leads to a minimized waiting time for
arriving passengers. If the descending traffic is light you cannot have an idle
car at every floor. Therefore minimizing the waiting time in light ascending
traffic is easier than minimizing it in light descending traffic.
According to Crites and Barto[6] it is the capacity for ascending traffic
that sets the capacity for the whole system. So when engineers try to build
a system of elevators, they should try to predict the heaviest up-peak in the
building and accommodate that demand. If the system is adjusted to the
heaviest up-peak, it will often be sufficient for down-peak as well.

Chapter 3
Method
The chosen method for this project is implementation of an elevator control
system since it gives a good overview of the system and it makes a reliable
source for results.

Elevator Simulator

The programming language chosen for the simulator was Java, because Java
is the language in which the authors are most proficient in and it suits this
kind of implementation well. Our simulation features two elevators in an
office building with ten floors.
The state variables of the system are presented in Figure 3.1. Object
refers to both elevators and passengers.
Each elevator car in the simulated building is designed to have their own
queue for requests, which consists of both request from the inside (made
by pressing buttons on the elevators control panel) and from the outside
(up/down button on each floor). The elevator is controlled by this queue
and always runs in the direction of the first destination put in the queue.
The elevator itself only adds requests to the queue made from the inside.
The other requests are handled by the building which has an overview of
what states the elevators are in.

Figure 3.1: The state variables of the elevator system

8

CHAPTER 3. METHOD

The queuing in the strategy is made based on a combination of states

of the elevators. It will look for an elevator that firstly goes in the same
direction as the passengers request and secondly still not yet have passed
the passengers position. The strategy also checks whether the elevator is full
or not. If none of the elevators fulfill these requirements the strategy will,
if possible, assign an idle elevator. If there are several elevators idling, the
closest one will be chosen.
If an idle elevator is not found, the passenger will be put in a waiting list.
It will later on be assigned to an elevators queue when an elevator has been
found.
The elevator itself also checks on certain states when stopping at a floor.
First of all, it decides if it should pick up or leave passengers. If the elevator
is empty, there is no one to leave. If there are passengers at the current floor,
there may be someone to pick up. When deciding which passenger to pick
up, the elevator checks that the passenger has got the same direction as the
elevator. If the elevator is full when it stops to pick up a passenger, that
passenger will be returned to the waiting list in the strategy.
The floors in the building have two main states, the up and down buttons.
These are used to check whether a request is already sent from the floor if a
new passenger arrives. The passenger has got a few states, of which most are
used by the strategy and elevator to be able to place the passengers request
in the queue.
The passengers are spawned and immediately given an arrival and destination floor, which depends on what time of the day is simulated at the
moment. To get a reasonable rate of arrival, we used an exponential distribution to generate the waiting times between generating new passengers.
An exponential distribution describes the time between events in a Poisson
process. A Poisson process is a process in which events, at a constant average
rate, occur independently and continuously[12].
One time step in our system includes spawning new passengers (if the
generated time is the same as the current time), then assign hall calls from
the waiting passengers and after that moving the elevators. An elevator
can move one floor per time unit. If the elevator is picking up or leaving
passengers, it will stand still for a total of six time units to simulate that
process.
To get a better understanding of our system, consider an example of
our elevator group control process at the moment a passenger arrives. The
passenger is positioned at the fifth floor with direction down and destination
9

CHAPTER 3. METHOD

lobby floor. The first elevator is at the second floor with direction up and
one request for floor nine in its queue. The second elevator is positioned
at floor seven with direction down and one passenger inside. It is destined
for the request in its queue, which is the lobby floor; the destination of the
passenger who is already inside the elevator. None of the elevators are full
and they both have closed their doors. There is no one else on the floor the
passenger arrived and none of the up- and down-buttons is pressed.
1. The passenger presses the hall call button at the fifth floor. The downbutton is pressed.
2. The building adds the request at floor five to the list of places waiting
to be served.
3. The strategy finds that the second elevator has the same direction
(down) as the passenger, its position is above the passengers current
position, the elevator is not full and it is not closing its doors at the
moment.
4. The strategy adds the new request to the second elevator. The queue
of the second elevator now contains floor five and the lobby floor. The
elevator begins to travel down.
5. The elevator travels down and eventually arrives at floor five. This
takes two time units.
6. The elevators position is now floor five. The passenger enters the elevator and the hall call button (down) is reset. The fifth floor is now empty
and the elevator contains one more passenger. The waiting time of the
elevator is increased, it has to stop and leave or pick up passengers.
Floor five is removed from the queue.
7. Since another passenger in the elevator already has requested to go to
the lobby floor, the elevators queue does not need to be changed.
8. The elevator waits for four time units and then begins to close its doors.
9. The elevator waits for two more time units and the doors are now
closed. This means that no more passengers on this floor can enter the
elevator.
10

CHAPTER 3. METHOD

10. The elevator finally begins to travel down to the lobby. This trip takes
five time units since it is five floors of travel.
11. The elevator arrives at the lobby, the passenger exits the elevator whose
queue is emptied of request. The elevators destination is now set to
idle.

Development of the First Strategies

As mentioned earlier, work began by implementing the basic collective control

strategy with some additions. When this strategy was complete, the zone
approach strategy was implemented. It is basically the same; however, in the
zone approach strategy the floors are split into separate zones where only one
elevator would serve its own zone.
Both of the strategies implemented are greedy and they do not do any
further predictions of what will happen after the requests in the current
direction. Imagine a situation where one elevator has got a request to drop
a passenger at floor nine, and is currently at eight, and the other elevator is
idle in the entrance. When using the basic strategy the idling elevator will
be chosen to handle a new request at floor seven even if the other elevator is
closer and will soon be empty. If the zone strategy is used instead, the new
passenger will be put on the waiting list and will later on be queued to the
elevator that is closer.

Simulation and Statistics

The system uses a discrete time scheme where one unit corresponds to three
seconds. The exponential distribution is used to simulate the arrival of 150
passengers during one hour.
The simulation was set to run for a time corresponding to two hours.
The building consists of 10 floors. Each passenger records their time when
entering an elevator and when leaving an elevator to get the waiting time
and the total time in system. When a round of the simulation is complete,
the calculated values are the average wait time, average total time and travel
times for the elevators compared to their passengers distances. The average
amount of passengers in the whole system is also calculated.

11

CHAPTER 3. METHOD

Optimized strategies

After running the simulation with the two strategies we compared them and
then added small optimizations to the most successful one.
We also choose to run the simulation a second round under heavier load
when we had optimized the strategy. We set the arrival rate to the double,
300 passengers during one hour, and ran the program for one hour instead,
to see what difference it made for the strategies.

4.1

First Version

The first small improvement we added to the basic strategy was to have a
default floor where the elevator is positioned while idle. This standard floor
is preferably the floor with the most incoming requests or in the case of the
collective control strategy during up-peak traffic, the lobby floor. In downpeak you could use the idea of the zone approach and have the two elevators
being idle at two different floors, one in the higher floors and one in the lower
floors so that the elevator will travel shorter distances to serve calls from the
whole building. We set the default floor of one elevator to floor two and the
other one to floor seven, unless it is during up-peak, then both of them are
set to the lobby floor.

4.2

Second Version

In the second optimization we improved the default floor feature. In the first
version, the movement to the default floor could not be interrupted which is
very ineffective when a new passenger arrives the second after the elevator
began to move to its default floor. Interruption of this movement was made
possible.
The cars were also given the possibility to change their queues during
special conditions. If an elevator has a request at a certain floor, and a new
passenger arrives one floor above that request, the elevator will change its
way to pick up the highest located passenger first if these passengers are
going in the same direction. This works the other way as well. In order to
implement this feature to be effective, a lot of states need to be checked to
avoid queuing the newly arrived passenger in the wrong situation.

12

Chapter 4
Results
Figure 4.1 shows our test results from simulations on the basic and the zone
strategy during regular pressure on the system. The time is presented in our
discrete time unit, which is three seconds. The average waiting time is the
time between the passengers arrival to the building until the boarding of an
elevator. The total time is the passengers whole time in the system. We also
measured how the passengers were distributed between the two elevators.

Initial Strategies

As can be seen in Figure 4.1 the zone strategy does not work at all in up-peak
traffic since all the passengers arrive in one zone which leaves one elevator
redundant. Furthermore, the results show that the basic strategy is more
efficient in both up-peak and down-peak, with respect to waiting time as
well as total time.

Optimization

Since the zone approach strategy gave unsatisfactory results even in downpeak we choose to optimize the collective control strategy, with some ideas
from the zone approach. The following comparisons are made only between
the basic strategy and the optimized versions.
The first optimization introduced a default floor for the elevators to move
to while idle, since the destination of the passengers never are the same as
the arrival floor in both up- and down-peak. By implementing the default
floor we saw a 17.6 percent decrease in average waiting time and a 9.1 percent
decrease in average total traveling time during up-peak traffic. During downpeak the waiting time was decreased with 10.9 percent and the average total
traveling time by 3.8 percent. To see the discrete values see Figure 4.2.
The system became even more flexible in the second iteration of the optimized strategy. The average waiting time for up-peak traffic decreased by
2.2 while the total time did not experience any significant decrease. During down-peak the waiting time decreased 27.0 percent while the total time
13

CHAPTER 4. RESULTS

Figure 4.1: Results from our Basic and Zone strategy

Figure 4.2: Results from test with the Basic strategy and the first optimization

Figure 4.3: Results from test with the first and the second optimization

14

CHAPTER 4. RESULTS

Figure 4.4: Average waiting time and total time during high pressure
decreased by 15.8 percent. These results indicate that the optimization was
successful and that by making the system as dynamic and flexible as possible
we gained a lot of performance. To see the discrete values for the average
waiting and total time, see Figure 4.3.

Higher Pressure

While running the simulation under a higher load (300 passengers during
one hour) the results were very different. The optimized version was slightly
better during up-peak and even better results for the waiting time and total
time during down-peak. Worth noticing is that the zone strategy got better
results than the basic strategy in one category; total time in down-peak.
Results are shown in Figure 4.4.

Energy Efficiency

We also chose to look at the ratio between the total distance the passengers
travelled and the distance the elevators travelled. This ratio tells us about
how energy efficient the strategy is. Figure 4.5 shows how much more the
elevators travel compared to the passengers. During down-peak the results
of the basic and the zone strategy were almost the same, but unfortunately
our optimizations showed disappointing results. The ratio was almost three
times worse, in both up- and down-peak.
During normal pressure, the zone strategy in up-peak is the only one
showing good results, however you cannot take that result in consideration

15

CHAPTER 4. RESULTS

Figure 4.5: Energy efficiency during normal pressure

since the zone strategy works like a single elevator system and the average
waiting and travel times are a lot worse than the others.

Source of Error

One source of error might be the fact that each test was not conducted on
the exact same amount of passengers since they were generated at random
times. We have tried to minimize this risk by testing all strategies 100 times
and then take the average results from these tests. Figure 4.6 shows the
difference in the amount of passengers who used the system during our test
times.
To simplify the logic and ease the testing of the system discrete units
were used instead of continuous time. This constitutes a source of error since
the value of our unit is approximated and not statistically determined.

16

CHAPTER 4. RESULTS

Figure 4.6: Average total amount of passengers in the system during different
strategies

17

Chapter 5
Discussion
We expected to see a gain in efficiency using the zone strategy during downpeak traffic, but the basic strategy was the best one in all cases. The only
time when the zone strategy actually had a better total time was under high
pressure in down-peak. However, the waiting time was still worse. This
seems reasonable since the travel time for the elevator in the upper zone will
always be longer, but under higher pressure it can at least pick up many
passengers at the same time.
The results were clear when it came to the basic strategy versus the
optimized strategy under regular pressure. The first optimizations increased
the performance a lot, especially during up-peak, the idle time was turned
into benefitting movements towards the next arrival.
The second round of optimization also gave interesting results. While
the first round of optimizations showed that our results for the up-peak were
slightly increased, this run showed no big change between the basic strategy
and the optimized one in up-peak. However, the times in down-peak was
once again decreased while testing the system with regular pressure.
When running the system under mheavier pressure, the optimizations
performance in the up-peak was just slightly better than the basic strategy.
Under higher pressure the default floor feature is not that useful and if the
elevator keep queuing requests one step over the current queue, the time
will be increased for the passengers that requested that elevator from the
beginning.
Moreover, it is known that people do not like to wait whether it is for the
elevator or the cashier. Minimizing the average waiting times is crucial for
passenger satisfaction and has more impact on the experience of using the
elevator than minimizing riding time[13]. This is because people make the
distinction between standing still on a floor and standing still in an elevator.
This is because a moving elevator feels like progression even though you are
not physically moving. Our optimizations reduced the average waiting time
by 4 percent in up-peak traffic and 7 percent in down-peak traffic. Thus
if this had been a real elevator system we would have had more satisfied
passengers after implementing the optimizations.
The energy efficiency results were not that surprising. Since both of the
18

CHAPTER 5. DISCUSSION

Figure 5.1: Average total traveling time divided into wait and travel
first strategies are pretty straightforward with the way they travel, we did
not expect anything special from them in terms of efficiency. However, the
fact that the optimized strategy increased the distance to such great numbers
was disappointing since it showed no other bad qualities. As expected, this
ratio was better during high pressure, since the elevators were able to pick
up more passengers along the way and use their full capacity.

19

Chapter 6
Conclusion
In summary, we have investigated two different elevator control strategies
for a simulated building with two elevators. The results show that there is
no need to change strategy between the up-peak and the down-peak traffic,
since one of the strategies was significantly better. However, a few small
optimizations could increase the performance of the system drastically. These
optimizations worsened the energy efficiency of the system, but instead both
the waiting time and the total time for the passengers were lowered, mainly
during down-peak.

20

Bibliography
[1] T. Takahashi and S. Matsuda, (2010), Adaptive elevator dispatching
with co-neuroevolution, Evolutionary Computation (CEC), 2010 IEEE
Congress on Print
[2] E4 project:
European Commissions Intelligent Energy Europe
Programme, (2010), Energy Efficient Elevators and Escalators,
http://www.e4project.eu/documenti/wp6/E4-WP6-Brochure.pdf,
March 2010, 090413
[3] Wlodzislaw Duch, (2007), What is Computational Intelligence and what
could it become?, http://cogprints.org/5358/1/06-CIdef.pdf, Jan 2007,
100413
[4] T. Beielstein, C. Ewald, S. Markon, (2003), Genetic and Evolutionary
Computation GECCO 2003, Springer Berlin Heidelberg, vol. 2724
[5] Marja-Liisa Siikonen, (1993), Elevator traffic simulation, SIMULATION,
vol.61
[6] Robert H. Crites, Andrew G. Barto, (1998), Elevator group control multiple using reinforcement learning agents, Machine Learning, vol. 33
[7] George R. Starkosch, Robert S. Caporale, (2010), The vertical transportation handbook, Wiley, fourth edition
[8] Chuen Chien Lee, (1990), Fuzzy logic in control systems: Fuzzy logic
controller- part 1, IEEE Transcations on systems. man. and cybernetics,
vol. 20
[9] Gu deying, Yan dongmei , (2010), Study on fuzzy algorithm of Elevator
Group Control System, Challenges in Environmental Science and Computer Engineering (CESCE), 2010 International Conference on Print, vol.
1
[10] Mat
Buckland,(2004),
AI-junkie,
junkie.com/ga/intro/gat1.html, July 2004, 100413

http://www.ai-

[11] M. Srinivas, Lalit M. Patnaik, (1994), Genetic Algorithms: A Survey,

Computer, vol. 27

21

BIBLIOGRAPHY

[12] Gunnar Blom, (2004), Sannolikhetsteori och statistikteori med

tillampningar, Studentlitteratur AB, fifth edition
[13] Jafferi Jamaludin, Nasrudin Abd. Rahim, Wooi Ping Hew, (2010), An
Elevator Group Control System With a Self-Tuning Fuzzy Logic Group
Controller, IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, vol. 57

22